基于改进Faster R-CNN的嘴部检测方法

在通过嘴部进行人机交互的场景下，外界光线变化、小目标检测的复杂性、检测方法的不通用性等因素给不同场景下嘴部的识别带来了很大困难.该文以不同场景下的人脸图像为数据源，提出了一种基于改进Faster R-CNN的人脸嘴部识别算法.该方法在Faster R-CNN框架中结合多尺度特征图进行检测，首先将同一卷积块不同卷积层输出的特征图结合，然后对不同的卷积块按元素进行求和操作，在输出的特征图上进行上采样得到高分辨率的表达能力更强的特征，从而提高了嘴部这种小目标的检测性能.在网络训练试验中运用多尺度训练和增加锚点数量增强网络检测不同尺寸目标的鲁棒性.实验表明，相比于原始的Faster R-CNN，对嘴部的检测准确率提高了8%，对环境的适应性更强.

Mouth Detection Method Based on Improved Faster R-CNN

In the scenario of human-computer interaction by the mouth, the light changes, the complexity of the small target detection, and the detection method of none generality factors under different scenarios have brought great difficulties to detect the mouth. In this study, we take the face images with different scenarios as data source and propose a face recognition algorithm based on Faster R-CNN. In this method, multi-scale feature maps are combined in Faster R-CNN framework for detection. Firstly, we introduce a modified multi-scale feature map to effectively utilize multi-resolution information. Then, feature maps need to share the same size, so that element-wise sum operation can be performed. Features with higher resolution and stronger expression ability can be obtained by up-sampling on the output feature map. The detection performance of the small target is improved. In the training experiment, multi-scale training and increasing the number of anchor points are used to enhance the robustness of the network to detect targets of different sizes. Experiments show that the detection accuracy of the mouth is improved by 8%, and it is more adaptable to the environment compared with the original Faster R-CNN.